The present invention relates to a speech pitch coding system for high quality coding of a speech signal at a low bit rate, particularly 4 kb/sec or lower.
A prior art speech coding system codes a speech signal based upon characteristic parameter data obtained for each frame (with a length of 40 msec., for instance) of the speech signal and characteristic parameter data obtained for each of sub-frames (with a length of 8 msec., for instance) as further divisions of the frame. The system comprises two excitation sources, i.e., an adaptive codebook produced by repeating a previous excitation signal at a pitch period and an excitation source codebook consisting of a previously produced signal, and produces a synthesized excitation signal by passing the excitation signal through a linear prediction synthesis filter. The synthesis filter is constructed using a filter coefficient set (for instance, a linear prediction filter coefficient set) obtained through analysis of a present frame input speech to be quantized. As such coding system, a CELP (Code-Excited LPC coding) system is well known, which is disclosed in, for instance, a treatise by M. Schroeder and B. Atal entitled "Code-Excited Linear Prediction: High Quality Speech at Very Low Bit Rates", IEEE Proc., ICASSP-85, pp. 937-940, 1985).
In another prior art system, the pitch coding in a small amount of operations by a pitch preliminary selection is performed. As such systems, there are a two-stage retrieval system (disclosed in Japanese Patent Laid-Open Publication No. Heisei 4-305135), which comprises steps of a pitch preliminary selection step in an open loop by using auto-correlation coefficients of a residual signal and a pitch final selection step from selected candidates by using a closed loop distortion, a two-stage retrieval system (disclosed in Japanese Patent Laid-Open No. Heisei 4-270398), which comprises steps of a pitch preliminary selection step in an open loop by using auto-correlation coefficients of an input signal and a final pitch selection step from delays close to selected candidates using a closed loop distortion, and a three-stage retrieval system (disclosed in TECHNICAL REPORT OF IEICE. SP92-133, 1993-02, Para. 5.1.2), which comprises steps of a preliminary pitch selection step in an open loop by using auto-correlation coefficients of a residual signal, a subsequent pitch preliminary selection step in a closed loop with sole inner product of an input signal and each codevector, and a pitch final selection step from selected candidates using a closed loop distortion.
In the above prior art systems, however, the pitch preliminary selection is performed in each sub-frame processing. Therefore, if the number of candidates in the pitch final selection is excessively reduced, a pitch with a locally small waveform distortion may be selected, increasing the speech quality deterioration of the coded speech. To avoid this problem, a certain number of candidates is required, thus making it difficult to reduce the amount of operations involved.